Skip to content

    Methodology

    How we score

    Exactly how the rankings are put together, and where they fall short.

    Last updated

    In short

    We rank AI using public data and a fixed method, and we show our working. Nobody can pay for a better position. Rankings are our opinion of how things compare. They are a starting point, not a guarantee that something will suit you.

    How we rank AI models

    Every model gets a score out of 10. The model at the top each week scores 10, and every other model is scored against it. The score is made of three parts:

    • What people prefer

      Share of the score
      About 52%
      Where it comes from
      Thousands of people compare two anonymous answers side by side and pick the better one (Arena)
    • Standard tests

      Share of the score
      About 43%
      Where it comes from
      Results on a wide set of standard tests, combined into one index (Epoch AI Capabilities Index)
    • Cost

      Share of the score
      5%
      Where it comes from
      The price developers pay to use the model. Cheaper scores higher

    Which models we include:

    • Only models the public can use today, checked on the maker's own website.
    • Only models released in the last two years.
    • "Free" means you can use it on the maker's free plan. Paying only for developer access doesn't count as free.
    • "Developers only" means you can't chat with it in the maker's app, only build it into software.
    • "Retiring" means the maker plans to switch it off within 90 days.
    • Models that don't yet have standard-test results are listed separately under "Not fully scored" and are not ranked. Scoring them on votes alone would flatter them.

    How sure we are: votes and tests always carry some uncertainty, so two models with close scores are, in practice, about as good as each other. Each week we re-run the ranking 5,000 times with small, realistic variations to check how settled each position is. You can see the result for any model under "How we measured it" in its details.

    "What it's good for" (writing, coding and so on) uses Arena's separate votes for each type of task. A model needs 9 out of 10 or more at a task to be called "among the best" at it.

    How we rank AI tools

    Tools are ranked within each category, such as "Make images" or "Meetings and notes". Tools are shown by rank only, without a score, because a score could suggest more precision than the data has.

    • Popularity-based rankings put tools in order of how many people use and rate them. We use Google's public Chrome UX Report for website reach, and the number of ratings on the App Store. A tool's star rating nudges its position up or down.
    • Tested rankings also show results from independent quality tests where they exist, such as blind comparisons of generated images.
    • Prices, free plans and availability come from each tool's own website or app store page. They are shown, but they never change a rank.
    • The UK and US versions use each country's own popularity and pricing data.
    • AI features built into a bigger product (for example a coding assistant inside an editor) are listed but not ranked against standalone tools.
    • Tools mainly for adult content are not ranked.

    How we choose news

    We score each possible story on five questions, from 0 to 2 each: does it change what an ordinary person can do or pay, does it involve a product lots of people use, does it touch safety, privacy, jobs, money or the law, are trusted outlets covering it, and is it a real first or milestone. Stories scoring 6 or more out of 10 are published. Stories scoring 8 or more are marked "Need to know".

    Every story needs an official source or two independent news outlets. We write each one in our own words and link to the original. The "Hype check" gives our honest view of whether the fuss matches the facts.

    How we write Learn guides

    Lessons and guides are written for people with no technical background, and checked against the makers' own help pages. We use AI tools to help research and draft, and every lesson and guide is reviewed by the haystak team before it goes live.

    Visitor ratings

    Anyone can rate an AI model or tool from 1 to 5 stars, and add a comment if they want to. These ratings show what other visitors think. They don't change our rankings or scores.

    Star ratings appear straight away. Comments appear only after we've read them, and you can report any comment you think is wrong or unfair. We don't record your name, email address or IP address when you rate.

    Update schedule

    The model rankings update every Monday morning (UK time). Tool rankings and prices are re-checked regularly. Every table shows the date it was last updated. If a data source is missing or looks wrong, we keep the previous ranking rather than publish a doubtful one.

    What we don't do

    • We don't accept payment for a ranking or a better position.
    • We don't invent or estimate numbers. If something is unknown, we say "not confirmed".
    • We don't let affiliate links or sponsors affect any rank.

    Limitations

    • Close scores can't be separated with confidence. Two models a few tenths apart are, in practice, about as good as each other.
    • Scores describe general use. A lower-ranked model or tool may be better for your particular task.
    • Prices and plans change often. Always check the maker's own website before you pay.
    • Popularity is not the same as quality. That's why we label which rankings are tested.

    Sources and credits

    "Adapted" means we have combined and rescaled the original data. The original publishers don't endorse haystak or our rankings.